WO2020031984A1 - Component inspection method and inspection system - Google Patents

Component inspection method and inspection system Download PDF

Info

Publication number
WO2020031984A1
WO2020031984A1 PCT/JP2019/030781 JP2019030781W WO2020031984A1 WO 2020031984 A1 WO2020031984 A1 WO 2020031984A1 JP 2019030781 W JP2019030781 W JP 2019030781W WO 2020031984 A1 WO2020031984 A1 WO 2020031984A1
Authority
WO
WIPO (PCT)
Prior art keywords
image data
inspection
model
input
noise
Prior art date
Application number
PCT/JP2019/030781
Other languages
French (fr)
Japanese (ja)
Inventor
希依彦 永井
Original Assignee
Ｂｌｕｅ Ｔａｇ株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ｂｌｕｅ Ｔａｇ株式会社 filed Critical Ｂｌｕｅ Ｔａｇ株式会社
Priority to JP2020535771A priority Critical patent/JPWO2020031984A1/en
Publication of WO2020031984A1 publication Critical patent/WO2020031984A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01B—MEASURING LENGTH, THICKNESS OR SIMILAR LINEAR DIMENSIONS; MEASURING ANGLES; MEASURING AREAS; MEASURING IRREGULARITIES OF SURFACES OR CONTOURS
    • G01B11/00—Measuring arrangements characterised by the use of optical techniques
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01B—MEASURING LENGTH, THICKNESS OR SIMILAR LINEAR DIMENSIONS; MEASURING ANGLES; MEASURING AREAS; MEASURING IRREGULARITIES OF SURFACES OR CONTOURS
    • G01B11/00—Measuring arrangements characterised by the use of optical techniques
    • G01B11/30—Measuring arrangements characterised by the use of optical techniques for measuring roughness or irregularity of surfaces
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N23/00—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00
    • G01N23/02—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00 by transmitting the radiation through the material
    • G01N23/04—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00 by transmitting the radiation through the material and forming images of the material
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis

Definitions

  • the present invention relates to a component inspection method, and more particularly, to a component inspection method and an inspection system for detecting an abnormality in an image of a component having substantially the same shape photographed from a predetermined position.
  • Conventional component inspection methods involve learning abnormal image data of a captured image, comparing the abnormal image data with the captured image data of the inspection target, and detecting an abnormality based on a correlation between the two. there were.
  • Patent Document 1 Japanese Patent Application Laid-Open No. 2007-279046, "Inspection System and Method” (Patent Document 1), and Japanese Patent Application Laid-Open No. 2010-139317, “Method and Apparatus for Defect Inspection of Tool Surface of Shaft” 2), JP-A-2013-205320, “Inspection condition determination method, inspection method and inspection device” (Patent Document 3), and JP-A-2017-054239, “Image classification device and image classification method” (Patent Document 4). is there.
  • Patent Literature 1 discloses a component inspection system that determines an actual parameter of a feature with respect to a reference point based on a captured video image and inspects the arrangement of the feature.
  • Patent Literature 2 discloses that an image obtained by photographing the surface of a shaft tool is divided into a plurality of pieces, and the presence or absence of a defect is determined based on an output value calculated by a learned neural network processing unit.
  • Patent Document 3 a typical defect image is learned by a classifier, the feature amount of the inspection target image is calculated, the inspection condition is determined by giving the characteristic amount to the classifier, and a defect is detected from the inspection target image. It has been shown.
  • Patent Literature 4 discloses that a plurality of attribute items are determined for a defect, a classifier is learned using a teacher data set, and a defect class is determined based on a result of classification of the plurality of attribute items for a target image. .
  • Patent Documents 1 to 4 it is difficult to detect various abnormal locations and abnormal aspects because the characteristic points of the abnormal locations are detected to determine the component abnormality.
  • the present invention has been made in view of the above situation, and provides a component inspection method and an inspection system that can easily detect abnormalities that appear irregularly in an image photographed from a fixed position with the shape of the inspection target component being constant.
  • the purpose is to do.
  • the present invention for solving the problems of the above conventional example is an inspection method in which an inspection of a component is performed by an inspection device using image data of a captured component. Is input as teacher data, model learning is performed by machine learning so that the output reconstructed image data has the same range as the input teacher data, and the trained model is generated as an inspection model, and the inspection model is generated. Enter the image data of the part photographed as the inspection target, compare the reconstructed image data with the input image data, and determine that both are normal if both are recognized as the same range. If not within the same range, it is determined to be abnormal.
  • noise-free image data is input as teacher data, and when noise-containing image data is input, machine learning is performed so that noise-free image data is output.
  • Model learning is performed, the learned model is generated as a noise removal model, the image data of the component photographed as the inspection target is input to the noise removal model, and the output noise-removed image data is used as the inspection model. What you enter.
  • the present invention in the inspection method, performs an operation of comparing the reconstructed image data and the input image data with a mean square error, and when the calculated value is within a specific threshold, the same range. If it is recognized and exceeds a certain threshold, it is not recognized as the same range.
  • the teacher data is divided into a plurality of small images to perform model learning, and the image data of a part photographed as an inspection target is also divided into a plurality of small images, and the divided image data is divided. Each time the data is input, the reconstructed image data is compared with the input image data.
  • the model learning is performed using an auto encoder or an AI model of a GAN.
  • the present invention is an inspection method for performing an inspection of a component using an inspection device using image data of a captured component.For image data of a component to be inspected, image data without noise is input as teacher data. Model learning by machine learning is performed so that noise-free image data is output when noise-free image data is input, the learned model is generated as a noise removal model, and the component photographed as an inspection target in the noise removal model The image data obtained by inputting the above image data and outputting the noise is estimated by comparing the LBP values of the image data using the LBP image processing to estimate the defect location.
  • the present invention relates to an inspection system for inspecting a part by using image data of an imaged part, a teacher data storage unit that stores image data of a normal part as teacher data, and a part photographed as an inspection target.
  • An image data storage unit for storing image data of the same, and inputting the teacher data from the teacher data storage unit, and performing model learning by machine learning so that the output reconstructed image data has the same range of contents as the input teacher data.
  • the learned model is generated as an inspection model, image data from the image data storage unit is input to the inspection model, and the reconstructed image data is compared with the input image data.
  • the inspection apparatus includes a control unit which determines that the range is normal, determines that the range is normal, and determines that both are not the same range, determines that the range is abnormal.
  • the teacher data storage unit stores, as teacher data, noise-free image data of the image data of the component to be inspected, and the control unit reads the noise data from the teacher data storage unit.
  • Input image data as teacher data perform model learning by machine learning so as to output image data without noise when image data with noise is input, generate the trained model as a noise removal model,
  • the image data of the component photographed as the inspection target from the image data storage unit is input to the removal model, and the output image data from which noise has been removed is input to the inspection model.
  • the present invention provides the inspection system, wherein the control unit performs an operation of comparing the reconstructed image data and the input image data with a mean square error, and the calculated value is within a specific threshold. Are recognized as the same range, and are not recognized as the same range when a certain threshold is exceeded.
  • the control unit performs model learning by dividing the teacher data into a plurality of small images, and also divides the image data of the part photographed as the inspection target into the plurality of small images. Each of the input image data is input, and the reconstructed image data is compared with the input image data.
  • an auto encoder or an AI model of a GAN is used for model learning.
  • the present invention relates to an inspection system for inspecting a component using image data of a captured component, wherein a teacher data storage unit stores image data without noise as teacher data for image data of a component to be inspected. And an image data storage unit for storing image data of a part photographed as an inspection target, and inputting image data without noise from the teacher data storage unit as teacher data. Model learning by machine learning is performed so that data is output, the learned model is generated as a noise removal model, image data of a part photographed as an inspection target is input to the noise removal model, and output noise is calculated. Deletion of the removed image data is performed by comparing the LBP values of the image data using LBP image processing. And it has a testing device and a control unit for estimating the Tokoro.
  • the inspection apparatus inputs image data of a normal part as teacher data, and performs model learning by machine learning so that the output reconstructed image data has the same range as the input teacher data.
  • Performing the learned model as an inspection model inputting image data of a part photographed as an inspection target to the inspection model, comparing the reconstructed image data with the input image data, Is determined to be normal if the same range is recognized, and abnormal if both are not recognized to be the same range. There is an effect that can be detected.
  • image data without noise is input as teacher data, and when image data with noise is input, model learning by machine learning is performed so that image data without noise is output.
  • Generating the trained model as a noise removal model inputting image data of a part photographed as an inspection target to the noise removal model, and inputting the output noise-removed image data to the inspection model. Since the inspection method is used, there is an effect that noise can be removed from image data before being input to the inspection model, and detection accuracy of an abnormal image can be improved.
  • FIG. 2 is a configuration block diagram of the present system. It is the schematic of AI utilization. It is a schematic diagram of an auto encoder.
  • FIG. 3 is a schematic diagram of image processing using an auto encoder. It is a flowchart of an inspection process. It is the schematic of the application example of this system.
  • the component inspection method according to the embodiment of the present invention uses a neural network to read normal image data as teacher data, performs machine learning using an auto encoder, and performs normal image data model ( Inspection model), and then input the image data for inspection into the inspection model and compare the reconstructed image data with the input inspection image data using an auto encoder.
  • Inspection model normal image data model
  • FIG. 1 is a configuration block diagram of the present system.
  • the present system includes an inspection device 1, a display unit 2, an input unit 3, and an image capturing device 4.
  • image data captured by the image capturing device 4 is taken into the inspection device 1 and processing for detecting an abnormality in the image data is performed.
  • the inspection device 1 includes a control unit 11, a storage device 12, and an interface unit 13.
  • the control unit 11 reads the processing program stored in the storage device 12 and executes an inspection process. Specific inspection processing will be described later.
  • the storage device 12 stores a processing program. Further, the storage device 12 stores an image data storage unit 121 that stores image data captured by the image capturing device 4, a teacher data storage unit 122 that stores normal image data as teacher data, and stores data of inspection results. And a test result data storage unit 123 to be executed.
  • the interface unit 13 is an interface for connecting to the display unit 2, the input unit 3, the image photographing device 4, and the network.
  • Display unit 2 displays image data captured by the image capturing device 4 and displays an inspection process and an inspection result.
  • the input unit 3 inputs a processing instruction in the present system.
  • the image photographing device 4 photographs a component from a plurality of locations at a fixed position, and outputs photographed image data to the inspection device 1.
  • the inspection device 4 for example, a device that photographs with X-rays is assumed.
  • the inspection component is a component having a uniform shape and a smooth surface. Note that the inspection content is to detect a defect such as a cavity inside the component or a fine scratch on the surface.
  • FIG. 2 is a schematic diagram of using AI. As shown in FIG. 2, the use of AI has a learning phase and an inference phase.
  • model learning is performed using the image data that has been inspected and determined to be normal as a learning data (teacher data) set to generate a learned model (inspection model).
  • learning data teacher data
  • learned model inspection model
  • machine learning is performed using an auto encoder, and normal image data can be correctly restored, but abnormal image data is learned so as not to be correctly restored.
  • image data taken for inspection is input as input data to a trained model (inspection model), and if the output data is within the same range as the input data, “normal” is determined. It is determined that it is "abnormal” if it is not within the same range.
  • an inspection model is generated by machine learning described below, and captured image data (input image data) of the inspection target is input to the inspection model, and the obtained output image data and The inspection is performed by comparing the input image data to determine whether the input image data is normal or abnormal.
  • the inspection apparatus 1 of the present system reads normal image data in advance as teacher data, and sends the normal image data to an auto encoder (Auto Encoder) which is a type of a deep learning neural network (DNN: Deep Neural Network).
  • Image data is input, subjected to image processing such as luminance mapping display, encoded, compressed and expressed, decoded (decompressed) to generate reproduced image data, and normal image data (input data) and reproduced image data. (Output data) to generate the same image data to generate (construct) a learned model (inspection model / AI model).
  • the normal image data is an entire image captured by the image capturing device 4, but may be a specific partial image portion of the image.
  • the inspection model uses a normal image group and utilizes the DNN of deep learning so as to extract the feature amount of a normal image, utilizing the fact that the image regions at the same position have similar properties. Build the encoder. Accordingly, if normal image data is input to the inspection model, the same normal image data is output. However, if abnormal image data (abnormal image data) is input, the same image data is output. It is something that was not done.
  • the inspection apparatus 1 inputs the captured image data (input image data) of the inspection target from the image data storage unit 121 to the inspection model having the above characteristics, and obtains the obtained output image data and input image data. Is determined, if the range is considered to be the same image data, the inspection component of the image is determined to be normal, and if the range is not determined to be the same image data, the inspection component of the image is determined to be abnormal. Things.
  • FIG. 3 is a schematic diagram of the auto encoder.
  • the original image data x is input to the input unit (Input)
  • the original image data x is encoded by a function f (x)
  • the encoded data h is converted to a function f (x).
  • Decoding is performed with g (h)
  • reconstructed (Reconstructed) image data x ′ is obtained from the output unit (Output).
  • the original image data and the reconstructed image data x ′ are compared, and if the difference between the two is equal to or smaller than a specific threshold value, the image is determined to be the same image (normal image). If the value exceeds a specific threshold value, it is determined that the image is another image (abnormal image).
  • the inspection model is generated by performing deep learning so that the reconstructed image data x ′ is determined to be the same.
  • FIG. 4 is a schematic diagram of image processing using an auto encoder.
  • an original image an image on which image processing such as luminance mapping display has been performed
  • Is output an image on which image processing such as luminance mapping display has been performed
  • the auto-encoder of the present system reads and learns a lot of normal original image data as teacher data, and if it is normal original image data, the difference value between the original image data and the reconstructed image data is both values. Deep learning is performed so that the images can be regarded as the same.
  • the difference between the input and output image data is within the range considered to be the same (the difference is equal to or less than a specific threshold), it is determined that the image data is normal and the range considered to be the same ( If the difference exceeds a certain threshold value and is not regarded as the same, it is determined that the image data is abnormal.
  • a reconstruction error value is calculated as the difference between the two images.
  • the calculated reconstruction error value is calculated using MSE (Mean Square Error: mean square error). If the error is within a specific threshold, the image is regarded as having the same image with a small error, and if the error exceeds a specific threshold. , An image having a large error is determined not to be regarded as the same image.
  • FIG. 5 is a flowchart of the inspection process.
  • the control unit 11 of the inspection apparatus 1 inputs image data captured as an inspection target to the inspection model (S1), and acquires image data output from the inspection model (S2).
  • the input image data and the output image data are compared (S3), and it is determined whether or not both are within the same range (S4). If it is determined that they are within the same range (Yes). It is determined that the image data is normal, and the "normal" inspection result is output to the inspection result data storage unit 123 and stored (S5).
  • noise is removed from the original image
  • the original image is divided into small area grids
  • model learning is performed using an auto encoder for each of the small areas
  • an inspection model generated by the model learning is used as an inspection target.
  • An abnormality detection process may be performed for each of the divided image data to specify an abnormal location in small area units.
  • FIG. 6 is a schematic diagram of an application example of the present system.
  • an AI model inspection model
  • image data for inspection is input to the AI model and output. If the image data and the input image data are not within the same range, the inspection image data is determined to be abnormal.
  • the apparatus includes a noise removing unit 21 for performing pre-removal processing, a defect image detecting unit 22 for detecting a defect image from image data from which noise has been removed, and a defect image detecting unit 23.
  • the noise removing unit 21, the defect image detecting unit 22, and the defect image detecting unit 23 in FIG. 6 are realized by the control unit 11 of the inspection apparatus 1 in FIG. 1 executing a processing program.
  • the auto encoder (Auto Encoder) 22a and the GAN (Generative Adversarial Network) 22b in the defect image detecting means 22 may be provided with both, or may be provided with either one. Further, both the defect image detection means 22 and the defect image detection means 23 may be provided, but either the defect image detection means 22 or the defect image detection means 23 may be provided.
  • noise removing means 21 Since the noise in the image data for inspection is similar to the defective portion, it is very important to remove the noise in order to improve the accuracy of the inspection in the present system. Specifically, normal two-dimensional image data (no-noise image data) in an ideal state is created using three-dimensional CAD (3D CAD: 3 Dimensional Computer Aided Design) data. The neural network (AI model / noise removal model) learns the two-dimensional image data as teacher data. The noise-free image data as the teacher data is stored in the teacher data storage unit 122 of the inspection apparatus 1 in FIG.
  • noise removal model is the noise removal means 21.
  • the noise removal model is constructed using, for example, a GAN (Generative Adversarial Network).
  • GAN Generic Adversarial Network
  • an AI algorithm other than GAN may be used.
  • GAN is referred to as a hostile generation network, is a type of AI algorithm used in unsupervised learning, and is implemented by a system of two neural networks that compete with each other in a zero-sum game framework.
  • the GAN is composed of two networks, a generator network and a discriminator network.
  • the generator of image generation outputs an image
  • the identifier determines whether the image is correct, and further, the generator deceives the identifier.
  • the discriminating side learns to try to discriminate more accurately.
  • the two networks are called "hostile" because they learn for conflicting purposes.
  • the noise removing unit 21 utilizes AI to remove only noise as image processing before detecting a defect because noise close to the defect is annoying.
  • AI is that, first, two-dimensional 3D CAD data can be used as ideal teacher data. Second, image processing is a smoothing filter applied to the entire image. This is because the AI can remove only those derived from noise during the learning process, without distinguishing the origins of defects.
  • the defect image detecting means 22 receives the image data from which noise has been removed from the noise removing means 21 and determines whether or not the input image data has a defect using the auto encoder 22a or the GAN 22b. The detection result of the image is output and stored in the inspection result data storage unit 123.
  • the processing in the auto encoder 22a is the same as that shown in FIGS. 1 to 5 in the present system described above.
  • the GAN 22b causes the GAN AI model (defective image detection model) to learn using normal image data as teacher data.
  • GAN AI model defective image detection model
  • an AI model that outputs the normal image data is constructed.
  • the defect image detecting means 22 divides the image data into an image of a small area of a fixed number of pixels (pixels), for example, 3 ⁇ 3 pixels, and performs a detection process for each of the small areas. Can be identified.
  • a dividing means for dividing the image into small areas is provided before inputting the image data to the auto encoder 22a or the GAN 22b.
  • the defect image detecting means 23 inputs the image data from which the noise has been removed from the noise removing means 21 and determines whether or not there is a defective portion in the input image data using an LBP (Local Binary Pattern) 23a. The result of detecting the abnormal image is output and stored in the inspection result data storage unit 123.
  • LBP Local Binary Pattern
  • the LBP 23a uses a 3 ⁇ 3 pixel unit feature amount (LBP value) calculated based on the relationship between the pixel values of the center pixel and the peripheral pixels, and is constant with the LBP values of other 3 ⁇ 3 pixels in the image data. As compared with the direction (for example, the lateral direction), the normal portion is overwhelmingly wide, and if the LBP value is significantly different from the surrounding LBP value, it is estimated that the portion is defective.
  • LBP value 3 ⁇ 3 pixel unit feature amount
  • the difference from the surrounding LBP value for estimating the defect location is determined, for example, by setting a threshold based on the average of the LPB values obtained from the image data, It is to judge.
  • an inspection system (another system) according to another embodiment
  • an AI model is trained using defect image data as teacher data, and when image data to be inspected is input, it is inferred whether or not the image data is defect image data.
  • a large amount of image data of a defect serving as teacher data is required.
  • the defect image data obtained by the present system is small, and another system cannot be realized.
  • defect pseudo image data having characteristics similar to defect image data is generated with many variations using AI.
  • GAN is used for generating defect pseudo image data. Specifically, features are extracted from real defect image data, and a large amount of defect pseudo image data having features common to the extracted features is generated while changing.
  • the neural network learns using the defect pseudo image data generated in large quantities as teacher data, and constructs an AI model (defect image detection model) for defect image detection.
  • the defect image detection model is assumed to be a CNN (Convolutional Neural Network).
  • CNN Convolutional Neural Network
  • the defect image detecting means 22 or the defect image detection 23 detects or detects and collects genuine defect image data, and when a specific amount of defect image data is collected, shifts to another system. Processing may be performed. That is, the system is operated until a specific amount of defective image data is accumulated, and thereafter, the system is switched to another system. Switching may be automated by program processing.
  • the present system and another system can coexist and can be used properly depending on the situation, and the accuracy of detection or detection can be improved.
  • the defect image detection means 22 and the defect image detection means 23 of FIG. 6 are provided, and furthermore, the defect image detection means 22 is provided with both the auto encoder 22a and the GAN 22b, and if both can be selected, Depending on the state of the image, the accuracy of defect detection or defect detection can be improved.
  • the inspection target is described as “parts”. However, not only parts of industrial products but also finished products obtained by combining a plurality of parts are included in parts. Shall be included.
  • the defect image can be detected with high accuracy.
  • the defect image is detected by the auto encoder 22a or the GAN 22b in the defect image detecting means 22, there is an effect that the defect image can be accurately detected depending on the state of the image data.
  • the defect image is detected by the LBP 23a in the defect image detecting means 23, there is an effect that the defect image can be accurately detected depending on the state of the image data.
  • the present invention is suitable for a component inspection method and an inspection system that can easily detect irregularly appearing abnormalities in an image photographed from a fixed position with a constant shape of the inspection target component.

Abstract

[Problem] To provide a component inspection method and a component inspection system with which it is possible to easily detect abnormalities that occur irregularly in images which are of to-be-inspected components having a definite shape and which are captured from a fixed position. [Solution] Provided are an inspection method and an inspection system, both comprising: by using a neural network, reading normal image data as teaching data so as to carry out machine learning on the data by using an autoencoder; constructing a model (inspection model) of the normal image data, from which only a normal image is to be accurately reconstructed; then, inputting a to-be-inspected image data set to said inspection model, and comparing, by using the autoencoder, the inputted to-be-inspected image data set with a reconstructed image data set; and determining the presence of an abnormality in the case where both image data sets are different to the extent of not being considered to be identical.

Description

部品の検査方法及び検査システムInspection method and inspection system for parts
　本発明は、部品の検査方法に係り、特に、ほぼ同じ形状の部品を定められた位置から撮影した画像について異常を検出する部品の検査方法及び検査システムに関する。 The present invention relates to a component inspection method, and more particularly, to a component inspection method and an inspection system for detecting an abnormality in an image of a component having substantially the same shape photographed from a predetermined position.
［従来の技術］
　従来の部品の検査方法には、撮影された画像の異常画像データを学習させて、その異常画像データと撮影された検査対象の画像データとを比較し、両者の相関によって異常を検出するものがあった。 [Conventional technology]
Conventional component inspection methods involve learning abnormal image data of a captured image, comparing the abnormal image data with the captured image data of the inspection target, and detecting an abnormality based on a correlation between the two. there were.
［関連技術］
　尚、関連する先行技術文献として、特開２００７－２７９０４６号公報「検査システムおよびその方法」（特許文献１）、特開２０１０－１３９３１７号公報「軸物工具表面の欠陥検査方法および装置」（特許文献２）、特開２０１３－２０５３２０号公報「検査条件決定方法、検査方法および検査装置」（特許文献３）、特開２０１７－０５４２３９号公報「画像分類装置および画像分類方法」（特許文献４）がある。 [Related technology]
As related prior art documents, Japanese Patent Application Laid-Open No. 2007-279046, "Inspection System and Method" (Patent Document 1), and Japanese Patent Application Laid-Open No. 2010-139317, "Method and Apparatus for Defect Inspection of Tool Surface of Shaft" 2), JP-A-2013-205320, “Inspection condition determination method, inspection method and inspection device” (Patent Document 3), and JP-A-2017-054239, “Image classification device and image classification method” (Patent Document 4). is there.
　特許文献１には、部品の検査システムであって、撮影したビデオ画像に基づいて参照点に対する特徴部の実際のパラメータを決定して特徴部の配置を検査することが示されている。
　特許文献２には、軸物工具の表面を撮影した画像を複数に分割し、学習されているニューラルネットワークの処理部が算出した出力値に基づいて欠陥の有無を判定することが示されている。 Patent Literature 1 discloses a component inspection system that determines an actual parameter of a feature with respect to a reference point based on a captured video image and inspects the arrangement of the feature.
Patent Literature 2 discloses that an image obtained by photographing the surface of a shaft tool is divided into a plurality of pieces, and the presence or absence of a defect is determined based on an output value calculated by a learned neural network processing unit.
　特許文献３には、典型的な欠陥画像を分類器に学習させておき、検査対象画像の特徴量を算出し、分類器に与えることで検査条件を決定し、検査対象画像から欠陥を検出することが示されている。
　特許文献４には、欠陥に対して複数の属性項目が定められ、分類器が教師データセットで学習され、対象画像に対する複数の属性項目の分類結果により欠陥クラスを決定することが示されている。 In Patent Document 3, a typical defect image is learned by a classifier, the feature amount of the inspection target image is calculated, the inspection condition is determined by giving the characteristic amount to the classifier, and a defect is detected from the inspection target image. It has been shown.
Patent Literature 4 discloses that a plurality of attribute items are determined for a defect, a classifier is learned using a teacher data set, and a defect class is determined based on a result of classification of the plurality of attribute items for a target image. .
特開２００７－２７９０４６号公報JP 2007-279046 A 特開２０１０－１３９３１７号公報JP 2010-139317 A 特開２０１３－２０５３２０号公報JP 2013-205320 A 特開２０１７－０５４２３９号公報JP 2017-054239 A
　しかしながら、上記従来の検査方法では、欠陥の形状等の異常がパターン化されたものであれば、異常検出に効果的であるが、欠陥の形状や発生箇所が一定ではなく異常が不規則に発生するものであれば、異常を容易に検出できないという問題点があった。 However, in the above-described conventional inspection method, if an abnormality such as the shape of a defect is patterned, it is effective for abnormality detection. However, the shape and location of the defect are not constant, and the abnormality occurs irregularly. However, there is a problem that the abnormality cannot be easily detected.
　特許文献１～４では、異常箇所の特徴点を検出して部品の異常を判定するため、様々な異常箇所、異常態様を検出するのが困難なものとなっていた。 (4) In Patent Documents 1 to 4, it is difficult to detect various abnormal locations and abnormal aspects because the characteristic points of the abnormal locations are detected to determine the component abnormality.
　本発明は上記実状に鑑みて為されたもので、検査対象部品の形状が一定で定位置から撮影される画像について、不規則に現れる異常を容易に検出できる部品の検査方法及び検査システムを提供することを目的とする。 The present invention has been made in view of the above situation, and provides a component inspection method and an inspection system that can easily detect abnormalities that appear irregularly in an image photographed from a fixed position with the shape of the inspection target component being constant. The purpose is to do.
（検査方法）
　上記従来例の問題点を解決するための本発明は、撮影された部品の画像データを用いて、部品の検査を検査装置で行う検査方法であって、検査装置では、正常な部品の画像データを教師データとして入力し、出力される再構成画像データが入力された教師データと同じ範囲の内容となるよう機械学習によるモデル学習を行い、当該学習済モデルを検査モデルとして生成し、当該検査モデルに検査対象として撮影された部品の画像データを入力し、再構成された画像データと入力された画像データとを比較して、両者が同じ範囲と認定される場合は正常と判定し、両者が同じ範囲と認定されない場合は異常と判定するものである。 (Inspection methods)
The present invention for solving the problems of the above conventional example is an inspection method in which an inspection of a component is performed by an inspection device using image data of a captured component. Is input as teacher data, model learning is performed by machine learning so that the output reconstructed image data has the same range as the input teacher data, and the trained model is generated as an inspection model, and the inspection model is generated. Enter the image data of the part photographed as the inspection target, compare the reconstructed image data with the input image data, and determine that both are normal if both are recognized as the same range. If not within the same range, it is determined to be abnormal.
（ノイズ除去）
　本発明は、上記検査方法において、検査対象の部品の画像データについて、ノイズのない画像データを教師データとして入力し、ノイズがある画像データを入力するとノイズのない画像データが出力されるよう機械学習によるモデル学習を行い、当該学習済モデルをノイズ除去モデルとして生成し、当該ノイズ除去モデルに検査対象として撮影された部品の画像データを入力し、出力されるノイズを除去した画像データを検査モデルに入力するものである。 (Noise removal)
According to the present invention, in the above-described inspection method, with respect to image data of a part to be inspected, noise-free image data is input as teacher data, and when noise-containing image data is input, machine learning is performed so that noise-free image data is output. Model learning is performed, the learned model is generated as a noise removal model, the image data of the component photographed as the inspection target is input to the noise removal model, and the output noise-removed image data is used as the inspection model. What you enter.
（平均二乗誤差で算出する検査方法）
　本発明は、上記検査方法において、再構成された画像データと入力された画像データを比較する演算を、平均二乗誤差で行い、当該演算された値が特定の閾値以内の場合には同じ範囲と認定され、特定の閾値を超える場合には同じ範囲と認定されないものである。 (Inspection method calculated by mean square error)
The present invention, in the inspection method, performs an operation of comparing the reconstructed image data and the input image data with a mean square error, and when the calculated value is within a specific threshold, the same range. If it is recognized and exceeds a certain threshold, it is not recognized as the same range.
（小地域分割の検査方法）
　本発明は、上記検査方法において、教師データを複数の小画像に分割してモデル学習を行い、検査対象として撮影された部品の画像データも複数の小画像に分割して、分割された画像データ毎に入力し、再構成された画像データと入力された画像データとの比較を行うものである。 (Inspection method of small area division)
According to the present invention, in the inspection method, the teacher data is divided into a plurality of small images to perform model learning, and the image data of a part photographed as an inspection target is also divided into a plurality of small images, and the divided image data is divided. Each time the data is input, the reconstructed image data is compared with the input image data.
（オートエンコーダ／ＧＡＮ）
　本発明は、上記検査方法において、モデル学習では、オートエンコーダ又はＧＡＮのＡＩモデルを用いて行うものである。 (Auto encoder / GAN)
According to the present invention, in the above inspection method, the model learning is performed using an auto encoder or an AI model of a GAN.
（ノイズ除去とＬＢＰの欠陥画像検出）
　本発明は、撮影された部品の画像データを用いて、部品の検査を検査装置で行う検査方法であって、検査対象の部品の画像データについて、ノイズのない画像データを教師データとして入力し、ノイズがある画像データを入力するとノイズのない画像データが出力されるよう機械学習によるモデル学習を行い、当該学習済モデルをノイズ除去モデルとして生成し、当該ノイズ除去モデルに検査対象として撮影された部品の画像データを入力し、出力されるノイズを除去した画像データについて、ＬＢＰの画像処理を用いて当該画像データにおけるＬＢＰ値の比較により欠陥箇所を推定するものである。 (Noise removal and LBP defect image detection)
The present invention is an inspection method for performing an inspection of a component using an inspection device using image data of a captured component.For image data of a component to be inspected, image data without noise is input as teacher data. Model learning by machine learning is performed so that noise-free image data is output when noise-free image data is input, the learned model is generated as a noise removal model, and the component photographed as an inspection target in the noise removal model The image data obtained by inputting the above image data and outputting the noise is estimated by comparing the LBP values of the image data using the LBP image processing to estimate the defect location.
（検査システム）
　本発明は、撮影され部品の画像データを用いて、部品の検査を行う検査システムであって、正常な部品の画像データを教師データとして記憶する教師データ記憶部と、検査対象として撮影された部品の画像データを記憶する画像データ記憶部と、教師データ記憶部から教師データを入力し、出力される再構成画像データが入力された教師データと同じ範囲の内容となるよう機械学習によるモデル学習を行い、当該学習済モデルを検査モデルとして生成し、当該検査モデルに画像データ記憶部からの画像データを入力し、再構成された画像データと入力された画像データとを比較して、両者が同じ範囲と認定される場合は正常と判定し、両者が同じ範囲と認定されない場合は異常と判定する制御部とを備える検査装置を有するものである。 (Inspection system)
The present invention relates to an inspection system for inspecting a part by using image data of an imaged part, a teacher data storage unit that stores image data of a normal part as teacher data, and a part photographed as an inspection target. An image data storage unit for storing image data of the same, and inputting the teacher data from the teacher data storage unit, and performing model learning by machine learning so that the output reconstructed image data has the same range of contents as the input teacher data. Then, the learned model is generated as an inspection model, image data from the image data storage unit is input to the inspection model, and the reconstructed image data is compared with the input image data. The inspection apparatus includes a control unit which determines that the range is normal, determines that the range is normal, and determines that both are not the same range, determines that the range is abnormal.
（ノイズ除去）
　本発明は、上記検査システムにおいて、教師データ記憶部には、検査対象の部品の画像データについて、ノイズのない画像データを教師データとして記憶しており、制御部が、教師データ記憶部からノイズのない画像データを教師データとして入力し、ノイズがある画像データを入力するとノイズのない画像データが出力されるよう機械学習によるモデル学習を行い、当該学習済モデルをノイズ除去モデルとして生成し、当該ノイズ除去モデルに画像データ記憶部からの検査対象として撮影された部品の画像データを入力し、出力されるノイズを除去した画像データを検査モデルに入力するものである。 (Noise removal)
According to the present invention, in the inspection system, the teacher data storage unit stores, as teacher data, noise-free image data of the image data of the component to be inspected, and the control unit reads the noise data from the teacher data storage unit. Input image data as teacher data, perform model learning by machine learning so as to output image data without noise when image data with noise is input, generate the trained model as a noise removal model, The image data of the component photographed as the inspection target from the image data storage unit is input to the removal model, and the output image data from which noise has been removed is input to the inspection model.
（平均二乗誤差で算出する検査システム）
　本発明は、上記検査システムにおいて、制御部が、再構成された画像データと入力された画像データを比較する演算を、平均二乗誤差で行い、当該演算された値が特定の閾値以内の場合には同じ範囲と認定され、特定の閾値を超える場合には同じ範囲と認定されないものである。 (Inspection system that calculates the mean square error)
The present invention provides the inspection system, wherein the control unit performs an operation of comparing the reconstructed image data and the input image data with a mean square error, and the calculated value is within a specific threshold. Are recognized as the same range, and are not recognized as the same range when a certain threshold is exceeded.
（小地域分割の検査システム）
　本発明は、上記検査システムにおいて、制御部が、教師データを複数の小画像に分割してモデル学習を行い、検査対象として撮影された部品の画像データも複数の小画像に分割して、分割された画像データ毎に入力し、再構成された画像データと入力された画像データとの比較を行うものである。 (Inspection system for small area division)
According to the present invention, in the inspection system, the control unit performs model learning by dividing the teacher data into a plurality of small images, and also divides the image data of the part photographed as the inspection target into the plurality of small images. Each of the input image data is input, and the reconstructed image data is compared with the input image data.
（オートエンコーダ／ＧＡＮ）
　本発明は、上記検査システムにおいて、モデル学習には、オートエンコーダ又はＧＡＮのＡＩモデルを用いたものである。 (Auto encoder / GAN)
According to the present invention, in the above inspection system, an auto encoder or an AI model of a GAN is used for model learning.
（ノイズ除去とＬＢＰの欠陥画像検出システム）
　本発明は、撮影された部品の画像データを用いて、部品の検査を行う検査システムであって、検査対象の部品の画像データについて、ノイズのない画像データを教師データとして記憶する教師データ記憶部と、検査対象として撮影された部品の画像データを記憶する画像データ記憶部と、教師データ記憶部からノイズのない画像データを教師データとして入力し、ノイズがある画像データを入力するとノイズのない画像データが出力されるよう機械学習によるモデル学習を行い、当該学習済モデルをノイズ除去モデルとして生成し、当該ノイズ除去モデルに検査対象として撮影された部品の画像データを入力し、出力されるノイズを除去した画像データについて、ＬＢＰの画像処理を用いて当該画像データにおけるＬＢＰ値の比較により欠陥箇所を推定する制御部とを備える検査装置を有するものである。 (Noise removal and LBP defect image detection system)
The present invention relates to an inspection system for inspecting a component using image data of a captured component, wherein a teacher data storage unit stores image data without noise as teacher data for image data of a component to be inspected. And an image data storage unit for storing image data of a part photographed as an inspection target, and inputting image data without noise from the teacher data storage unit as teacher data. Model learning by machine learning is performed so that data is output, the learned model is generated as a noise removal model, image data of a part photographed as an inspection target is input to the noise removal model, and output noise is calculated. Deletion of the removed image data is performed by comparing the LBP values of the image data using LBP image processing. And it has a testing device and a control unit for estimating the Tokoro.
　本発明によれば、検査装置が、正常な部品の画像データを教師データとして入力し、出力される再構成画像データが入力された教師データと同じ範囲の内容となるよう機械学習によるモデル学習を行い、当該学習済モデルを検査モデルとして生成し、当該検査モデルに検査対象として撮影された部品の画像データを入力し、再構成された画像データと入力された画像データとを比較して、両者が同じ範囲と認定される場合は正常と判定し、両者が同じ範囲と認定されない場合は異常と判定する検査方法としているので、定位置で撮影される画像データについて不規則に現れる異常を容易に検出できる効果がある。 According to the present invention, the inspection apparatus inputs image data of a normal part as teacher data, and performs model learning by machine learning so that the output reconstructed image data has the same range as the input teacher data. Performing the learned model as an inspection model, inputting image data of a part photographed as an inspection target to the inspection model, comparing the reconstructed image data with the input image data, Is determined to be normal if the same range is recognized, and abnormal if both are not recognized to be the same range. There is an effect that can be detected.
　本発明によれば、検査対象の部品の画像データについて、ノイズのない画像データを教師データとして入力し、ノイズがある画像データを入力するとノイズのない画像データが出力されるよう機械学習によるモデル学習を行い、当該学習済モデルをノイズ除去モデルとして生成し、当該ノイズ除去モデルに検査対象として撮影された部品の画像データを入力し、出力されるノイズを除去した画像データを検査モデルに入力する上記検査方法としているので、検査モデルに入力される前の画像データからノイズを除去でき、異常な画像の検出精度を向上させることができる効果がある。 According to the present invention, for image data of a component to be inspected, image data without noise is input as teacher data, and when image data with noise is input, model learning by machine learning is performed so that image data without noise is output. Generating the trained model as a noise removal model, inputting image data of a part photographed as an inspection target to the noise removal model, and inputting the output noise-removed image data to the inspection model. Since the inspection method is used, there is an effect that noise can be removed from image data before being input to the inspection model, and detection accuracy of an abnormal image can be improved.
本システムの構成ブロック図である。FIG. 2 is a configuration block diagram of the present system. ＡＩ利用の概略図である。It is the schematic of AI utilization. オートエンコーダの概略図である。It is a schematic diagram of an auto encoder. オートエンコーダを用いた画像処理の概略図である。FIG. 3 is a schematic diagram of image processing using an auto encoder. 検査処理のフローチャートである。It is a flowchart of an inspection process. 本システムの応用例の概略図である。It is the schematic of the application example of this system.
　本発明の実施の形態について図面を参照しながら説明する。
［実施の形態の概要］
　本発明の実施の形態に係る部品の検査方法（本検査方法）は、ニューラルネットワークを用いて、正常な画像データを教師データとして読み込んでオートエンコーダを用いて機械学習させて正常画像データのモデル（検査モデル）を構築し、その後に検査用の画像データを当該検査モデルに入力してオートエンコーダを用いて、再構成された画像データと入力された検査用の画像データとを比較して、両者が同一と見なされる範囲を超えていれば、異常と判定するものであり、定位置で撮影される画像データについて不規則に現れる異常を容易に検出できるものである。 An embodiment of the present invention will be described with reference to the drawings.
[Summary of Embodiment]
The component inspection method according to the embodiment of the present invention (the inspection method) uses a neural network to read normal image data as teacher data, performs machine learning using an auto encoder, and performs normal image data model ( Inspection model), and then input the image data for inspection into the inspection model and compare the reconstructed image data with the input inspection image data using an auto encoder. Are determined to be abnormal if the values are outside the range considered to be the same, and abnormalities that appear irregularly in image data captured at a fixed position can be easily detected.
［本システム：図１］
　本発明の実施の形態に係る検査システム（本システム）について図１を参照しながら説明する。図１は、本システムの構成ブロック図である。
　本システムは、図１に示すように、検査装置１と、表示部２と、入力部３と、画像撮影装置４とを備えている。
　本システムでは、画像撮影装置４で撮影された画像データを検査装置１に取り込み、画像データの異常を検出する処理を行うものである。 [This system: Fig. 1]
An inspection system (this system) according to an embodiment of the present invention will be described with reference to FIG. FIG. 1 is a configuration block diagram of the present system.
As shown in FIG. 1, the present system includes an inspection device 1, a display unit 2, an input unit 3, and an image capturing device 4.
In the present system, image data captured by the image capturing device 4 is taken into the inspection device 1 and processing for detecting an abnormality in the image data is performed.
［本システムの各部］
　本システムの各部について具体的に説明する。
　［検査装置１］
　検査装置１は、制御部１１と、記憶装置１２と、インタフェース部１３とを備えている。
　制御部１１は、記憶装置１２に記憶された処理プログラムを読み込んで検査処理を実行する。具体的な検査処理は後述する。 [Each part of the system]
Each part of the system will be specifically described.
[Inspection device 1]
The inspection device 1 includes a control unit 11, a storage device 12, and an interface unit 13.
The control unit 11 reads the processing program stored in the storage device 12 and executes an inspection process. Specific inspection processing will be described later.
　記憶装置１２は、処理プログラムを記憶する。
　また、記憶装置１２は、画像撮影装置４で撮影された画像データを記憶する画像データ記憶部１２１と、正常な画像データを教師データとして記憶する教師データ記憶部１２２と、検査結果のデータを記憶する検査結果データ記憶部１２３とを備えている。 The storage device 12 stores a processing program.
Further, the storage device 12 stores an image data storage unit 121 that stores image data captured by the image capturing device 4, a teacher data storage unit 122 that stores normal image data as teacher data, and stores data of inspection results. And a test result data storage unit 123 to be executed.
　インタフェース部１３は、表示部２、入力部３、画像撮影装置４、更にネットワークに接続するインタフェースである。 The interface unit 13 is an interface for connecting to the display unit 2, the input unit 3, the image photographing device 4, and the network.
　［表示部２，入力部３］
　表示部２は、画像撮影装置４で撮影された画像データを表示し、検査過程、検査結果を表示する。
　入力部３は、本システムにおける処理の指示を入力する。 [Display unit 2, Input unit 3]
The display unit 2 displays image data captured by the image capturing device 4 and displays an inspection process and an inspection result.
The input unit 3 inputs a processing instruction in the present system.
　［画像撮影装置４］
　画像撮影装置４は、部品を定位置で複数箇所から撮影し、撮影した画像データを検査装置１に出力する。画像撮影装置４として、例えば、Ｘ線で撮影する装置を想定している。また、検査用の部品としては、形状が一定で表面が滑らかな部品を想定している。尚、検査内容としては、部品内部の空洞等の欠陥や表面の微細なキズ等を検出するものである。 [Image capturing device 4]
The image photographing device 4 photographs a component from a plurality of locations at a fixed position, and outputs photographed image data to the inspection device 1. As the image photographing device 4, for example, a device that photographs with X-rays is assumed. Also, it is assumed that the inspection component is a component having a uniform shape and a smooth surface. Note that the inspection content is to detect a defect such as a cavity inside the component or a fine scratch on the surface.
［ＡＩ利用の概略：図２］
　次に、検査装置１におけるＡＩ利用の概略について図２を参照しながら説明する。図２は、ＡＩ利用の概略図である。
　図２に示すように、ＡＩ利用では、学習フェーズと推論フェーズがある。 [Outline of AI use: Fig. 2]
Next, the outline of the use of AI in the inspection apparatus 1 will be described with reference to FIG. FIG. 2 is a schematic diagram of using AI.
As shown in FIG. 2, the use of AI has a learning phase and an inference phase.
　学習フェーズで、検査済の正常と判定された画像データを学習用データ（教師データ）セットとして用いてモデル学習を行い、学習済モデル（検査モデル）を生成する。このモデル学習では、オートエンコーダを用いて機械学習を行って、正常な画像データは正しく復元できるが、異常な画像データは正しく復元できないように学習させる。 (4) In the learning phase, model learning is performed using the image data that has been inspected and determined to be normal as a learning data (teacher data) set to generate a learned model (inspection model). In this model learning, machine learning is performed using an auto encoder, and normal image data can be correctly restored, but abnormal image data is learned so as not to be correctly restored.
　そして、推論フェーズで、検査用に撮影された画像データを入力データとして、学習済モデル（検査モデル）に入力して、その出力データが入力データと同一の範囲内にあれば、「正常」と判定し、同一の範囲内になければ「異常」と判定するものである。 In the inference phase, image data taken for inspection is input as input data to a trained model (inspection model), and if the output data is within the same range as the input data, “normal” is determined. It is determined that it is "abnormal" if it is not within the same range.
［処理概略］
　次に、本システムにおける処理動作の概略を説明する。
　本システムの検査装置１では、以下に説明する機械学習によって検査モデルを生成し、検査対象の撮影された画像データ（入力画像データ）をその検査モデルに入力して、得られた出力画像データと入力画像データを比較して、入力画像データの正常／異常を判定して検査を行うものである。 [Process outline]
Next, an outline of a processing operation in the present system will be described.
In the inspection device 1 of the present system, an inspection model is generated by machine learning described below, and captured image data (input image data) of the inspection target is input to the inspection model, and the obtained output image data and The inspection is performed by comparing the input image data to determine whether the input image data is normal or abnormal.
　具体的には、本システムの検査装置１は、正常な画像データを教師データとして予め読み込んで、深層学習のニューラルネットワーク（ＤＮＮ：Deep Neural Network）の一種であるオートエンコーダ（Auto Encoder）に正常な画像データを入力して、輝度マッピング表示等の画像処理を行って符号化し、圧縮表現し、復号化（復元）して再生画像データを生成し、正常な画像データ（入力データ）と再生画像データ（出力データ）とが同じ画像データとなるよう学習させて学習済モデル（検査モデル／ＡＩモデル）を生成（構築）する。
　ここで、正常な画像データは、画像撮影装置４で撮影された１枚の画像全体としているが、その画像の特定の一部の画像部分を対象にしてもよい。 Specifically, the inspection apparatus 1 of the present system reads normal image data in advance as teacher data, and sends the normal image data to an auto encoder (Auto Encoder) which is a type of a deep learning neural network (DNN: Deep Neural Network). Image data is input, subjected to image processing such as luminance mapping display, encoded, compressed and expressed, decoded (decompressed) to generate reproduced image data, and normal image data (input data) and reproduced image data. (Output data) to generate the same image data to generate (construct) a learned model (inspection model / AI model).
Here, the normal image data is an entire image captured by the image capturing device 4, but may be a specific partial image portion of the image.
　当該検査モデルは、正常な画像群を用いて、同じ位置にある画像の領域は性質が似ていることを利用して、正常な画像の特徴量を取り出すように深層学習のＤＮＮを用いてオートエンコーダを構築する。
　これにより、当該検査モデルに正常な画像データが入力されれば、同じ正常な画像データが出力されるが、正常ではない画像データ（異常な画像データ）が入力されると、同じ画像データは出力されないようにしたものである。 The inspection model uses a normal image group and utilizes the DNN of deep learning so as to extract the feature amount of a normal image, utilizing the fact that the image regions at the same position have similar properties. Build the encoder.
Accordingly, if normal image data is input to the inspection model, the same normal image data is output. However, if abnormal image data (abnormal image data) is input, the same image data is output. It is something that was not done.
　そして、検査装置１は、上記の特徴を備えた検査モデルに、検査対象の撮影された画像データ（入力画像データ）を画像データ記憶部１２１から入力し、得られた出力画像データと入力画像データとを比較し、同一の画像データとみなされる範囲であれば、当該画像の検査部品を正常と判定し、同一の画像データとみなされる範囲でなければ、当該画像の検査部品を異常と判定するものである。 Then, the inspection apparatus 1 inputs the captured image data (input image data) of the inspection target from the image data storage unit 121 to the inspection model having the above characteristics, and obtains the obtained output image data and input image data. Is determined, if the range is considered to be the same image data, the inspection component of the image is determined to be normal, and if the range is not determined to be the same image data, the inspection component of the image is determined to be abnormal. Things.
　［オートエンコーダの概略：図３］
　次に、オートエンコーダについて図３を参照しながら説明する。図３は、オートエンコーダの概略図である。
　図３に示すように、オリジナル（Original）の画像データｘが入力部（Input）に入力されると、関数ｆ（ｘ）で符号化（Encode）し、符号化された符号化データｈを関数ｇ（ｈ）で復号化（Decode）して出力部（Output）から再構成（Reconstruct）画像データｘ´を得る。 [Outline of auto encoder: Fig. 3]
Next, the auto encoder will be described with reference to FIG. FIG. 3 is a schematic diagram of the auto encoder.
As shown in FIG. 3, when the original image data x is input to the input unit (Input), the original image data x is encoded by a function f (x), and the encoded data h is converted to a function f (x). Decoding is performed with g (h), and reconstructed (Reconstructed) image data x ′ is obtained from the output unit (Output).
　本システムの検査装置１では、オリジナルの画像データと再構成画像データｘ´を比較して、両者の差分の値が特定の閾値以下であれば、同じ画像（正常な画像）と判定し、差分の値が特定の閾値を超えていれば、別の画像（異常な画像）と判定する。
　尚、本システムのオートエンコーダでは、正常なオリジナルの画像データであれば、再構成画像データｘ´が同じと判定されるよう深層学習させて検査モデルを生成している。 In the inspection apparatus 1 of the present system, the original image data and the reconstructed image data x ′ are compared, and if the difference between the two is equal to or smaller than a specific threshold value, the image is determined to be the same image (normal image). If the value exceeds a specific threshold value, it is determined that the image is another image (abnormal image).
In the auto encoder of the present system, if the original image data is normal, the inspection model is generated by performing deep learning so that the reconstructed image data x ′ is determined to be the same.
　［オートエンコーダを用いた画像処理の概略：図４］
　次に、オートエンコーダを用いた画像処理について図４を参照しながら説明する。図４は、オートエンコーダを用いた画像処理の概略図である。
　本システムの検査装置１におけるオートエンコーダでは、図４に示すように、オリジナル画像（輝度マッピング表示等の画像処理が為された画像）を符号化して圧縮表現し、それを復元して再構成画像を出力する。 [Outline of image processing using auto encoder: FIG. 4]
Next, image processing using the auto encoder will be described with reference to FIG. FIG. 4 is a schematic diagram of image processing using an auto encoder.
In the auto-encoder in the inspection apparatus 1 of the present system, as shown in FIG. 4, an original image (an image on which image processing such as luminance mapping display has been performed) is encoded and expressed in a compressed manner. Is output.
　ここで、本システムのオートエンコーダは、正常なオリジナル画像データを教師データとして数多く読み込んで学習させ、正常なオリジナル画像データであれば、オリジナル画像データと再構成画像データとの差分の値が、両画像を同じと見なすことができる範囲となるよう深層学習されている。 Here, the auto-encoder of the present system reads and learns a lot of normal original image data as teacher data, and if it is normal original image data, the difference value between the original image data and the reconstructed image data is both values. Deep learning is performed so that the images can be regarded as the same.
　つまり、本システムの検査モデルでは、入力と出力の画像データの差分が同一と見なされる範囲内（差分が特定の閾値以下）であれば、正常な画像データと判定し、同一と見なされる範囲（差分が特定の閾値）を超えて同一と見なされないものになると、異常な画像データと判定するものである。 That is, in the inspection model of the present system, if the difference between the input and output image data is within the range considered to be the same (the difference is equal to or less than a specific threshold), it is determined that the image data is normal and the range considered to be the same ( If the difference exceeds a certain threshold value and is not regarded as the same, it is determined that the image data is abnormal.
　オリジナル画像と再構成画像との違いを定量的に把握するために、両画像の相違として再構成エラー値を算出する。演算される再構成エラー値は、ＭＳＥ（Mean Square Error：平均二乗誤差）を用いて算出し、特定の閾値以内である場合に、誤差が少なく同一の画像とみなし、特定の閾値を超える場合に、誤差が大きく同一の画像とはみなさない判定を行うものである。 (4) In order to quantitatively grasp the difference between the original image and the reconstructed image, a reconstruction error value is calculated as the difference between the two images. The calculated reconstruction error value is calculated using MSE (Mean Square Error: mean square error). If the error is within a specific threshold, the image is regarded as having the same image with a small error, and if the error exceeds a specific threshold. , An image having a large error is determined not to be regarded as the same image.
　［検査対象の画像データの検査処理：図５］
　次に、検査対象の画像データの検査処理について図５を参照しながら説明する。図５は、検査処理のフローチャートである。
　検査装置１の制御部１１が、当該検査モデルに検査対象として撮影した画像データを入力し（Ｓ１）、検査モデルから出力された画像データを取得する（Ｓ２）。 [Inspection processing of image data to be inspected: FIG. 5]
Next, an inspection process of image data to be inspected will be described with reference to FIG. FIG. 5 is a flowchart of the inspection process.
The control unit 11 of the inspection apparatus 1 inputs image data captured as an inspection target to the inspection model (S1), and acquires image data output from the inspection model (S2).
　そして、入力された画像データと出力される画像データとを比較し（Ｓ３）、両者が同一の範囲内かどうかを判定し（Ｓ４）、同一の範囲内と判定されれば（Ｙｅｓの場合）、正常な画像データであると判定し、「正常」の検査結果を検査結果データ記憶部１２３に出力して記憶させる（Ｓ５）。 Then, the input image data and the output image data are compared (S3), and it is determined whether or not both are within the same range (S4). If it is determined that they are within the same range (Yes). It is determined that the image data is normal, and the "normal" inspection result is output to the inspection result data storage unit 123 and stored (S5).
　また、同一の範囲内と判定されなければ（Ｎｏの場合）、異常な画像データであると判定し、「異常」の検査結果を検査結果データ記憶部１２３に出力して記憶させ（Ｓ６）、処理を終了する。 If it is not determined that the image data is within the same range (in the case of No), it is determined that the image data is abnormal, and the inspection result of “abnormal” is output to and stored in the inspection result data storage unit 123 (S6). The process ends.
　［異常個所の検出］
　以上の検査方法では、異常がある画像データを判別できるものではあるが、その異常のある画像データのどの場所（位置）に異常があるのかまでは特定していない。
　異常箇所を特定するために、異常箇所が判明しているサンプルの異常な画像データを画像データ記憶部１２１に記憶し、検査対象の画像データで異常と判定された画像データとサンプルの画像データとの相関を演算し、相関の高いサンプル画像データから異常箇所を推定するようにしてもよい。例えば、何番目のピクセルが異常の原因となっているかを突き止めることができる。 [Detection of abnormal part]
In the above inspection method, although image data having an abnormality can be determined, the location (position) of the image data having the abnormality is not specified.
In order to identify the abnormal part, abnormal image data of the sample in which the abnormal part is known is stored in the image data storage unit 121, and the image data determined to be abnormal in the image data to be inspected and the image data of the sample are stored. May be calculated, and an abnormal portion may be estimated from sample image data having a high correlation. For example, it is possible to determine what pixel is causing the abnormality.
　また、オリジナル画像からノイズを除去し、オリジナル画像を小領域のグリッドに分割し、その小領域毎にオートエンコーダを用いてモデル学習を行い、当該モデル学習で生成された検査モデルで、検査対象として分割された画像データ毎に異常検出処理を行って異常箇所を小地域単位で特定するようにしてもよい。 Also, noise is removed from the original image, the original image is divided into small area grids, model learning is performed using an auto encoder for each of the small areas, and an inspection model generated by the model learning is used as an inspection target. An abnormality detection process may be performed for each of the divided image data to specify an abnormal location in small area units.
［応用例：図６］
　本システムの応用例１について図６を参照しながら説明する。図６は、本システムの応用例の概略図である。
　本システムでは、正常な画像データを教師データとしてオートエンコーダで学習させ、正常な画像データを再構成させるＡＩモデル（検査モデル）を構築し、検査用の画像データをそのＡＩモデルに入力し、出力画像データと入力画像データが同一の範囲内になければ、検査用の画像データを異常と判断するものであるが、応用例では、図６に示すように、検査用の画像データに対してノイズ除去の前処理を行うノイズ除去手段２１と、ノイズ除去された画像データから欠陥画像を検知する欠陥画像検知手段２２と、欠陥画像検出手段２３とを備えている。 [Application example: Fig. 6]
An application example 1 of the present system will be described with reference to FIG. FIG. 6 is a schematic diagram of an application example of the present system.
In this system, an AI model (inspection model) for reconstructing normal image data is constructed by learning normal image data as teacher data by an auto encoder, and image data for inspection is input to the AI model and output. If the image data and the input image data are not within the same range, the inspection image data is determined to be abnormal. However, in the application example, as shown in FIG. The apparatus includes a noise removing unit 21 for performing pre-removal processing, a defect image detecting unit 22 for detecting a defect image from image data from which noise has been removed, and a defect image detecting unit 23.
　図６におけるノイズ除去手段２１、欠陥画像検知手段２２と欠陥画像検出手段２３は、図１の検査装置１の制御部１１が処理プログラムを実行することで実現される。
　また、欠陥画像検知手段２２におけるオートエンコーダ（Auto Encoder）２２ａ、ＧＡＮ（Generative Adversarial Network）２２ｂは、両方を備えていてもよいが、いずれか一つを備えているものであってもよい。
　また、欠陥画像検知手段２２と欠陥画像検出手段２３の双方を備えていてもよいが、欠陥画像検知手段２２又は欠陥画像検出手段２３のいずれかを備えるようにしてもよい。 The noise removing unit 21, the defect image detecting unit 22, and the defect image detecting unit 23 in FIG. 6 are realized by the control unit 11 of the inspection apparatus 1 in FIG. 1 executing a processing program.
The auto encoder (Auto Encoder) 22a and the GAN (Generative Adversarial Network) 22b in the defect image detecting means 22 may be provided with both, or may be provided with either one.
Further, both the defect image detection means 22 and the defect image detection means 23 may be provided, but either the defect image detection means 22 or the defect image detection means 23 may be provided.
　［ノイズ除去手段２１］
　検査用の画像データにおけるノイズと欠陥部が似ているため、ノイズを除去することは、本システムにおける検査の精度を向上させる上で非常に重要である。
　具体的には、３次元ＣＡＤ（３ＤＣＡＤ：3 Dimensional Computer Aided Design）データを用いて理想状態の正常な２次元画像データ（ノイズなし画像データ）を作成する。
　この２次元画像データを教師データとしてニューラルネットワーク（ＡＩモデル／ノイズ除去モデル）に学習させる。
　教師データとしてのノイズなし画像データは、図１の検査装置１の教師データ記憶部１２２に記憶させておく。 [Noise removing means 21]
Since the noise in the image data for inspection is similar to the defective portion, it is very important to remove the noise in order to improve the accuracy of the inspection in the present system.
Specifically, normal two-dimensional image data (no-noise image data) in an ideal state is created using three-dimensional CAD (3D CAD: 3 Dimensional Computer Aided Design) data.
The neural network (AI model / noise removal model) learns the two-dimensional image data as teacher data.
The noise-free image data as the teacher data is stored in the teacher data storage unit 122 of the inspection apparatus 1 in FIG.
　そして、ノイズありの２次元画像データを用意して画像データ記憶部１２１に記憶させておき、それらノイズありの画像データを使用して、ノイズありの画像データを入力すると、ノイズなしの画像データが出力されるノイズ除去モデルを構築する。
　このノイズ除去モデルがノイズ除去手段２１である。
　これにより、検査用の画像データをノイズ除去モデル（ノイズ除去手段２１）に入力すると、ノイズが除去された出力画像データが得られ、ノイズ除去された画像データを欠陥画像検知手段２２に入力する。 Then, two-dimensional image data with noise is prepared and stored in the image data storage unit 121, and the image data with noise is input using the image data with noise. Construct a noise removal model to be output.
This noise removal model is the noise removal means 21.
Thus, when the image data for inspection is input to the noise elimination model (noise elimination means 21), output image data from which noise has been eliminated is obtained, and the image data from which noise has been eliminated is input to the defect image detection means 22.
　ノイズ除去モデルには、例えば、ＧＡＮ（Generative Adversarial Network）を用いて構築する。尚、ＧＡＮ以外のＡＩアルゴリズムを用いても構わない。
　ＧＡＮは、敵対的生成ネットワークと言われ、教師なし学習で使用されるＡＩアルゴリズムの一種であり、ゼロサムゲームフレームワークで互いに競合する２つのニューラルネットワークのシステムによって実装される。 The noise removal model is constructed using, for example, a GAN (Generative Adversarial Network). Note that an AI algorithm other than GAN may be used.
GAN is referred to as a hostile generation network, is a type of AI algorithm used in unsupervised learning, and is implemented by a system of two neural networks that compete with each other in a zero-sum game framework.
　ＧＡＮは、生成ネットワーク（generator）と識別ネットワーク（discriminator）の２つのネットワークから構成され、画像生成の生成側がイメージを出力し、識別側がその正否を判定し、更に、生成側は識別側を欺くように学習し、識別側はより正確に識別しようと学習する。このように２つのネットワークが相反した目的で学習するため「敵対的」と称される。 The GAN is composed of two networks, a generator network and a discriminator network. The generator of image generation outputs an image, the identifier determines whether the image is correct, and further, the generator deceives the identifier. And the discriminating side learns to try to discriminate more accurately. Thus, the two networks are called "hostile" because they learn for conflicting purposes.
　ノイズ除去手段２１は、欠陥に近いノイズが厄介な存在であるため、欠陥を検知する前の画像処理として、ノイズだけを除去するのにＡＩを活用したものである。
　ＡＩを利用する理由は、第１に、３ＤＣＡＤデータを２次元化したものを理想的な教師データとして活用でき、第２に、画像処理は画像全体に掛ける平滑化フィルタであるため、画像由来と欠陥由来を区別しないが、ＡＩは学習過程でノイズ由来のものだけを取り除けるためである。 The noise removing unit 21 utilizes AI to remove only noise as image processing before detecting a defect because noise close to the defect is annoying.
The reason for using AI is that, first, two-dimensional 3D CAD data can be used as ideal teacher data. Second, image processing is a smoothing filter applied to the entire image. This is because the AI can remove only those derived from noise during the learning process, without distinguishing the origins of defects.
　［欠陥画像検知手段２２］
　欠陥画像検知手段２２は、ノイズ除去手段２１からのノイズを除去された画像データを入力し、オートエンコーダ２２ａ又はＧＡＮ２２ｂを用いて入力画像データに欠陥があるのか否かを判定し、正常／異常の画像を検知した結果を出力し、検査結果データ記憶部１２３に記憶する。
　オートエンコーダ２２ａでの処理は、上述した本システムにおける図１～５に示したものとなる。 [Defect image detecting means 22]
The defect image detecting means 22 receives the image data from which noise has been removed from the noise removing means 21 and determines whether or not the input image data has a defect using the auto encoder 22a or the GAN 22b. The detection result of the image is output and stored in the inspection result data storage unit 123.
The processing in the auto encoder 22a is the same as that shown in FIGS. 1 to 5 in the present system described above.
　ＧＡＮ２２ｂは、具体的には、正常な画像データを教師データとしてＧＡＮのＡＩモデル（欠陥画像検知モデル）に学習させる。
　そして、異常な画像データを使用して、それら異常な画像データを入力すると、正常な画像データが出力されるＡＩモデル（ＧＡＮ２２ｂ）を構築する。 Specifically, the GAN 22b causes the GAN AI model (defective image detection model) to learn using normal image data as teacher data.
When the abnormal image data is input using the abnormal image data, an AI model (GAN 22b) that outputs the normal image data is constructed.
　これにより、ノイズが除去された画像データを欠陥画像検知手段２２のＧＡＮ２２ｂに入力すると、正常な出力画像データが得られ、入力画像データと出力画像データとを比較し、両データが同一の範囲内か否かを判定し、同一の範囲内であれば正常な画像データと判定し、同一の範囲内でなければ、異常な画像データと判定するものである。
　同一の範囲内か否かの判定手法は、本システムの図４で説明したものと同様である。 Thus, when the image data from which noise has been removed is input to the GAN 22b of the defect image detecting means 22, normal output image data is obtained, and the input image data and the output image data are compared, and both data are within the same range. It is determined whether or not the image data is normal, if it is within the same range, and is determined as abnormal image data if it is not within the same range.
The method of determining whether the values are within the same range is the same as that described in FIG. 4 of the present system.
　また、欠陥画像検知手段２２は、画像データを一定数の画素（ピクセル）、例えば、３×３のピクセルの小領域の画像に分割し、その小領域毎に検知処理を行えば、異常箇所を特定できる。
　この場合、欠陥画像検知手段２２内で、オートエンコーダ２２ａ又はＧＡＮ２２ｂに画像データを入力する前段に小領域の画像に分割する分割手段を設けるものとする。 Further, the defect image detecting means 22 divides the image data into an image of a small area of a fixed number of pixels (pixels), for example, 3 × 3 pixels, and performs a detection process for each of the small areas. Can be identified.
In this case, in the defect image detecting means 22, a dividing means for dividing the image into small areas is provided before inputting the image data to the auto encoder 22a or the GAN 22b.
　［欠陥画像検出手段２３］
　欠陥画像検出手段２３は、ノイズ除去手段２１からのノイズを除去された画像データを入力し、ＬＢＰ（Local Binary Pattern）２３ａを用いて入力画像データにおける欠陥部位があるのか否かを判定し、正常／異常の画像を検出した結果を出力し、検査結果データ記憶部１２３に記憶する。 [Defect image detecting means 23]
The defect image detecting means 23 inputs the image data from which the noise has been removed from the noise removing means 21 and determines whether or not there is a defective portion in the input image data using an LBP (Local Binary Pattern) 23a. The result of detecting the abnormal image is output and stored in the inspection result data storage unit 123.
　ＬＢＰ２３ａは、中心画素と周辺画素の画素値の関係性を元に算出される３×３の画素単位の特徴量（ＬＢＰ値）を用い、画像データにおける他の３×３画素のＬＢＰ値と一定方向（例えば、横方向）に比較していき、正常部位が圧倒的に広範囲であることから、周辺のＬＢＰ値と大きく異なる場合、欠陥箇所と推定するものである。 The LBP 23a uses a 3 × 3 pixel unit feature amount (LBP value) calculated based on the relationship between the pixel values of the center pixel and the peripheral pixels, and is constant with the LBP values of other 3 × 3 pixels in the image data. As compared with the direction (for example, the lateral direction), the normal portion is overwhelmingly wide, and if the LBP value is significantly different from the surrounding LBP value, it is estimated that the portion is defective.
　欠陥箇所と推定するための周囲のＬＢＰ値との相違は、例えば、画像データで得られるＬＰＢ値の平均等を基にしきい値を設定し、当該しきい値を超えた場合に、欠陥箇所と判定するものである。 The difference from the surrounding LBP value for estimating the defect location is determined, for example, by setting a threshold based on the average of the LPB values obtained from the image data, It is to judge.
［別のシステム］
　次に、別の実施の形態に係る検査システム（別のシステム）について説明する。
　別のシステムでは、欠陥画像データを教師データとしてＡＩモデルに学習させ、検査対象の画像データを入力すると、欠陥画像データか否かを推論するものである。
　そのためには、教師データとなる欠陥の画像データを大量に必要になるが、本システムで得られる欠陥画像データは少なく、別のシステムを実現できない。 [Another system]
Next, an inspection system (another system) according to another embodiment will be described.
In another system, an AI model is trained using defect image data as teacher data, and when image data to be inspected is input, it is inferred whether or not the image data is defect image data.
For that purpose, a large amount of image data of a defect serving as teacher data is required. However, the defect image data obtained by the present system is small, and another system cannot be realized.
　そこで、別のシステムでは、ＡＩを用いて欠陥画像データに似た特徴を持つ「欠陥擬似画像データ」を多くのバリエーションをもって生成する。
　例えば、欠陥擬似画像データの生成に、ＧＡＮを用いる。具体的には、本物の欠陥画像データから特徴を抽出し、抽出した特徴に共通する特徴を有する欠陥擬似画像データを変化させながら大量に生成する。 Therefore, in another system, “defect pseudo image data” having characteristics similar to defect image data is generated with many variations using AI.
For example, GAN is used for generating defect pseudo image data. Specifically, features are extracted from real defect image data, and a large amount of defect pseudo image data having features common to the extracted features is generated while changing.
　そして、大量に生成した欠陥擬似画像データを教師データとして、ニューラルネットワークに学習させ、欠陥画像検出のＡＩモデル（欠陥画像検出モデル）を構築する。
　この欠陥画像検出モデルには、ＣＮＮ（Convolutional Neural Network：畳み込みニューラルネットワーク）を想定しており、学習済みモデルに検査対象の画像データを入力すると、画像データにおける欠陥箇所を探知するものである。 Then, the neural network learns using the defect pseudo image data generated in large quantities as teacher data, and constructs an AI model (defect image detection model) for defect image detection.
The defect image detection model is assumed to be a CNN (Convolutional Neural Network). When image data to be inspected is input to a learned model, a defect location in the image data is detected.
　別のシステムでは、欠陥擬似画像データを生成するためには、本物の欠陥画像データを収集する必要がある。本システムとの関係では、欠陥画像検知手段２２又は欠陥画像検出２３で本物の欠陥画像データを検知又は検出して収集し、特定の量の欠陥画像データが収集されたら、別システムに移行して処理を行うことが考えられる。
　つまり、特定の量の欠陥画像データが蓄積されるまでは、本システムを動作し、その後、別のシステムに切り替えるようにする。切り替えは、プログラム処理により自動化してもよい。 In another system, it is necessary to collect genuine defect image data in order to generate pseudo defect image data. In relation to the present system, the defect image detecting means 22 or the defect image detection 23 detects or detects and collects genuine defect image data, and when a specific amount of defect image data is collected, shifts to another system. Processing may be performed.
That is, the system is operated until a specific amount of defective image data is accumulated, and thereafter, the system is switched to another system. Switching may be automated by program processing.
　また、本システムと別のシステムを併存させ、状況によって使い分け、検知又は検出の精度を向上させることができる。この場合、図６の欠陥画像検知手段２２と欠陥画像検出手段２３の両方を備え、更に、欠陥画像検知手段２２では、オートエンコーダ２２ａとＧＡＮ２２ｂの双方を備えて、いずれも選択可能とすれば、画像の状況によって欠陥検知又は欠陥検出の精度を向上させることができる。 Also, the present system and another system can coexist and can be used properly depending on the situation, and the accuracy of detection or detection can be improved. In this case, if both the defect image detection means 22 and the defect image detection means 23 of FIG. 6 are provided, and furthermore, the defect image detection means 22 is provided with both the auto encoder 22a and the GAN 22b, and if both can be selected, Depending on the state of the image, the accuracy of defect detection or defect detection can be improved.
　尚、本システム及び別のシステムでは、検査対象を「部品」として説明したが、工業製品の部品に限らず、複数の部品の組み合わせた完成品も部品に含まれ、更に、食品等も部品に含まれるものとする。 In this system and another system, the inspection target is described as “parts”. However, not only parts of industrial products but also finished products obtained by combining a plurality of parts are included in parts. Shall be included.
［実施の形態の効果］
　本システム及び本方法によれば、ニューラルネットワークを用いて、正常な画像データを教師データとして読み込んでオートエンコーダを用いて機械学習させ、正常な画像データのみが正確に再構成される正常画像データのモデル（検査モデル）を構築し、その後に検査用の画像データを当該検査モデルに入力してオートエンコーダを用いて、再構成された画像データと入力された検査用の画像データとを比較して、両者が同一と見なされる範囲を超えていれば、異常と判定するものであり、定位置で撮影される画像データについて不規則に現れる異常を容易に検出できる効果がある。 [Effects of Embodiment]
According to the present system and method, normal image data is read as teacher data using a neural network, machine-learned using an auto encoder, and only normal image data is reconstructed correctly. A model (inspection model) is constructed, and the image data for inspection is input to the inspection model, and the reconstructed image data is compared with the input image data for inspection using an auto encoder. If both are out of the range considered to be the same, it is determined that there is an abnormality, and there is an effect that irregularities appearing irregularly in image data photographed at a fixed position can be easily detected.
　本システムの応用例によれば、欠陥画像を検知する前処理として、ノイズ除去手段２１で画像データにおけるノイズを除去しているので、欠陥画像の検知を精度よく行うことができる効果がある。 According to the application example of the present system, since noise in the image data is removed by the noise removing unit 21 as preprocessing for detecting a defect image, the defect image can be detected with high accuracy.
　また、本システムの応用例では、欠陥画像検知手段２２におけるオートエンコーダ２２ａ又はＧＡＮ２２ｂで欠陥画像の検知を行うため、画像データの状況によって欠陥画像の検知を精度よく行うことができる効果がある。 In addition, in the application example of the present system, since the defect image is detected by the auto encoder 22a or the GAN 22b in the defect image detecting means 22, there is an effect that the defect image can be accurately detected depending on the state of the image data.
　更に、本システムの応用例によれば、欠陥画像検出手段２３におけるＬＢＰ２３ａで欠陥画像の検出を行うため、画像データの状況によって欠陥画像の検出を精度よく行うことができる効果がある。 According to the application example of the present system, since the defect image is detected by the LBP 23a in the defect image detecting means 23, there is an effect that the defect image can be accurately detected depending on the state of the image data.
　本発明は、検査対象部品の形状が一定で定位置から撮影される画像について、不規則に現れる異常を容易に検出できる部品の検査方法及び検査システムに好適である。 The present invention is suitable for a component inspection method and an inspection system that can easily detect irregularly appearing abnormalities in an image photographed from a fixed position with a constant shape of the inspection target component.
　１…検査装置、　２…表示部、　３…入力部、　４…画像撮影装置、　１１…制御部、　１２…記憶装置、　１３…インタフェース部、　２１…ノイズ除去手段、　２２…欠陥画像検知手段、　２２ａ…オートエンコーダ、　２２ｂ…ＧＡＮ、　２３…欠陥画像検出手段、　２３ａ…ＬＢＰ、　１２１…画像データ記憶部、　１２２…教師データ記憶部、　１２３…検査結果データ記憶部 DESCRIPTION OF SYMBOLS 1 ... Inspection apparatus, # 2 ... Display part, # 3 ... Input part, # 4 ... Image photographing apparatus, # 11 ... Control part, # 12 ... Storage device, # 13 ... Interface part, # 21 ... Noise removal means, # 22 ... Defect image detection means, # 22a ... Auto encoder, # 22b GAN, # 23 defective image detecting means, # 23a LBP, # 121 image data storage unit, # 122 teacher data storage unit, # 123 inspection result data storage unit

Claims (12)

  1. 　撮影された部品の画像データを用いて、部品の検査を検査装置で行う検査方法であって、
    　前記検査装置では、正常な部品の画像データを教師データとして入力し、出力される再構成画像データが前記入力された教師データと同じ範囲の内容となるよう機械学習によるモデル学習を行い、当該学習済モデルを検査モデルとして生成し、当該検査モデルに検査対象として撮影された部品の画像データを入力し、再構成された画像データと前記入力された画像データとを比較して、両者が同じ範囲と認定される場合は正常と判定し、前記両者が同じ範囲と認定されない場合は異常と判定する検査方法。     An inspection method for inspecting a component with an inspection device using image data of a captured component,
    In the inspection device, image data of a normal part is input as teacher data, and model learning is performed by machine learning so that output reconstructed image data has the same range as the input teacher data. Generated model as an inspection model, image data of a part photographed as an inspection target is input to the inspection model, the reconstructed image data is compared with the input image data, and both are in the same range. Inspection method in which the normal range is determined when the determination is made, and the abnormality is determined when the two ranges are not determined to be in the same range.
  2. 　検査対象の部品の画像データについて、ノイズのない画像データを教師データとして入力し、ノイズがある画像データを入力するとノイズのない画像データが出力されるよう機械学習によるモデル学習を行い、当該学習済モデルをノイズ除去モデルとして生成し、当該ノイズ除去モデルに検査対象として撮影された部品の画像データを入力し、出力されるノイズを除去した画像データを検査モデルに入力する請求項１記載の検査方法。 With respect to the image data of the component to be inspected, input image data without noise as teacher data, perform model learning by machine learning so that when image data with noise is input, image data without noise is output. 2. The inspection method according to claim 1, wherein the model is generated as a noise elimination model, image data of a part photographed as an inspection target is input to the noise elimination model, and image data from which noise is output is input to the inspection model. .
  3. 　再構成された画像データと入力された画像データを比較する演算を、平均二乗誤差で行い、当該演算された値が特定の閾値以内の場合には同じ範囲と認定され、前記特定の閾値を超える場合には同じ範囲と認定されない請求項１又は２記載の検査方法。 The operation of comparing the reconstructed image data and the input image data is performed with a mean square error, and when the calculated value is within a specific threshold, it is recognized as being in the same range and exceeds the specific threshold. The inspection method according to claim 1 or 2, wherein the case is not recognized as being in the same range.
  4. 　教師データを複数の小画像に分割してモデル学習を行い、検査対象として撮影された部品の画像データも前記複数の小画像に分割して、前記分割された画像データ毎に入力し、再構成された画像データと前記入力された画像データとの比較を行う請求項１乃至３のいずれか記載の検査方法。 The teacher data is divided into a plurality of small images, model learning is performed, and image data of a part photographed as an inspection target is also divided into the plurality of small images, and input is performed for each of the divided image data, and reconstruction is performed. The inspection method according to claim 1, wherein the comparison is performed between the input image data and the input image data.
  5. 　モデル学習では、オートエンコーダ又はＧＡＮのＡＩモデルを用いて行う請求項１乃至４のいずれか記載の検査方法。 The inspection method according to any one of claims 1 to 4, wherein the model learning is performed using an auto encoder or an AI model of a GAN.
  6. 　撮影された部品の画像データを用いて、部品の検査を検査装置で行う検査方法であって、
    　検査対象の部品の画像データについて、ノイズのない画像データを教師データとして入力し、ノイズがある画像データを入力するとノイズのない画像データが出力されるよう機械学習によるモデル学習を行い、当該学習済モデルをノイズ除去モデルとして生成し、当該ノイズ除去モデルに検査対象として撮影された部品の画像データを入力し、出力されるノイズを除去した画像データについて、ＬＢＰの画像処理を用いて当該画像データにおけるＬＢＰ値の比較により欠陥箇所を推定する検査方法。     An inspection method for inspecting a component with an inspection device using image data of a captured component,
    With respect to the image data of the component to be inspected, input image data without noise as teacher data, perform model learning by machine learning so that when image data with noise is input, image data without noise is output. A model is generated as a noise removal model, image data of a part photographed as an inspection target is input to the noise removal model, and the output noise-removed image data is processed using LBP image processing. An inspection method for estimating a defective portion by comparing LBP values.
  7. 　撮影され部品の画像データを用いて、部品の検査を行う検査システムであって、
    　正常な部品の画像データを教師データとして記憶する教師データ記憶部と、
    　検査対象として撮影された部品の画像データを記憶する画像データ記憶部と、
    　前記教師データ記憶部から教師データを入力し、出力される再構成画像データが前記入力された教師データと同じ範囲の内容となるよう機械学習によるモデル学習を行い、当該学習済モデルを検査モデルとして生成し、当該検査モデルに前記画像データ記憶部からの画像データを入力し、再構成された画像データと前記入力された画像データとを比較して、両者が同じ範囲と認定される場合は正常と判定し、前記両者が同じ範囲と認定されない場合は異常と判定する制御部とを備える検査装置を有する検査システム。     An inspection system for inspecting a part using image data of the part being photographed,
    A teacher data storage unit that stores image data of normal parts as teacher data,
    An image data storage unit that stores image data of a part photographed as an inspection target,
    Inputting teacher data from the teacher data storage unit, performing model learning by machine learning so that reconstructed image data to be output has the same range of contents as the input teacher data, and using the learned model as an inspection model Generate and input the image data from the image data storage unit to the inspection model, compare the reconstructed image data with the input image data. And a control unit that determines that there is an abnormality when the two are not recognized as being in the same range.
  8. 　教師データ記憶部には、検査対象の部品の画像データについて、ノイズのない画像データを教師データとして記憶しており、
    　制御部は、前記教師データ記憶部からノイズのない画像データを教師データとして入力し、ノイズがある画像データを入力するとノイズのない画像データが出力されるよう機械学習によるモデル学習を行い、当該学習済モデルをノイズ除去モデルとして生成し、当該ノイズ除去モデルに画像データ記憶部からの検査対象として撮影された部品の画像データを入力し、出力されるノイズを除去した画像データを検査モデルに入力する請求項７記載の検査システム。     The teacher data storage unit stores noise-free image data as teacher data for the image data of the component to be inspected.
    The control unit inputs image data without noise as the teacher data from the teacher data storage unit, performs model learning by machine learning such that when image data with noise is input, image data without noise is output, Generated as a noise removal model, image data of the part photographed as an inspection target from the image data storage unit is input to the noise removal model, and the output noise-removed image data is input to the inspection model. The inspection system according to claim 7.
  9. 　制御部は、再構成された画像データと入力された画像データを比較する演算を、平均二乗誤差で行い、当該演算された値が特定の閾値以内の場合には同じ範囲と認定され、前記特定の閾値を超える場合には同じ範囲と認定されない請求項７又は８記載の検査システム。 The control unit performs an operation of comparing the reconstructed image data and the input image data with a mean square error, and when the calculated value is within a specific threshold value, it is determined that the calculated value is within the same range. The inspection system according to claim 7 or 8, wherein the same range is not recognized when the threshold value is exceeded.
  10. 　制御部は、教師データを複数の小画像に分割してモデル学習を行い、検査対象として撮影された部品の画像データも前記複数の小画像に分割して、前記分割された画像データ毎に入力し、再構成された画像データと前記入力された画像データとの比較を行う請求項７乃至９のいずれか記載の検査システム。 The control unit divides the teacher data into a plurality of small images, performs model learning, divides the image data of the part photographed as the inspection target into the plurality of small images, and inputs the image data for each of the divided image data. The inspection system according to claim 7, wherein a comparison is made between the reconstructed image data and the input image data.
  11. 　モデル学習には、オートエンコーダ又はＧＡＮのＡＩモデルを用いた請求項７乃至１０のいずれか記載の検査システム。 The inspection system according to any one of claims 7 to 10, wherein in the model learning, an auto encoder or an AI model of GAN is used.
  12. 　撮影された部品の画像データを用いて、部品の検査を行う検査システムであって、
    　検査対象の部品の画像データについて、ノイズのない画像データを教師データとして記憶する教師データ記憶部と、
    　検査対象として撮影された部品の画像データを記憶する画像データ記憶部と、
    　前記教師データ記憶部からノイズのない画像データを教師データとして入力し、ノイズがある画像データを入力するとノイズのない画像データが出力されるよう機械学習によるモデル学習を行い、当該学習済モデルをノイズ除去モデルとして生成し、当該ノイズ除去モデルに検査対象として撮影された部品の画像データを入力し、出力されるノイズを除去した画像データについて、ＬＢＰの画像処理を用いて当該画像データにおけるＬＢＰ値の比較により欠陥箇所を推定する制御部とを備える検査装置を有する検査システム。     An inspection system for inspecting a component by using image data of a captured component,
    For image data of a component to be inspected, a teacher data storage unit that stores image data without noise as teacher data;
    An image data storage unit that stores image data of a part photographed as an inspection target,
    Inputting image data without noise as the teacher data from the teacher data storage unit, performing model learning by machine learning such that when image data with noise is input, image data without noise is output. The image data of the component which is generated as a removal model and which is photographed as an inspection target is input to the noise removal model, and the LBP value of the image data from which the output noise is removed is determined using LBP image processing. An inspection system including an inspection device including a control unit that estimates a defect location by comparison.
PCT/JP2019/030781 2018-08-08 2019-08-05 Component inspection method and inspection system WO2020031984A1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP2020535771A JPWO2020031984A1 (en) 2018-08-08 2019-08-05 Parts inspection method and inspection system

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2018149153 2018-08-08
JP2018-149153 2018-08-08

Publications (1)

Publication Number Publication Date
WO2020031984A1 true WO2020031984A1 (en) 2020-02-13

Family

ID=69414776

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2019/030781 WO2020031984A1 (en) 2018-08-08 2019-08-05 Component inspection method and inspection system

Country Status (2)

Country Link
JP (1) JPWO2020031984A1 (en)
WO (1) WO2020031984A1 (en)

Cited By (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021106566A1 (en) * 2019-11-25 2021-06-03 オムロン株式会社 Image inspection device
JP2021140739A (en) * 2020-02-28 2021-09-16 株式会社Pros Cons Program, method for creating learned model, information processing method and information processing device
JP2021157847A (en) * 2020-11-16 2021-10-07 ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッドBeijing Baidu Netcom Science Technology Co., Ltd. Method, apparatus, device, and readable storage medium for recognizing abnormal license plate
WO2021210612A1 (en) * 2020-04-16 2021-10-21 浜松ホトニクス株式会社 Radiographic image acquiring device, radiographic image acquiring system, and radiographic image acquisition method
WO2021210618A1 (en) * 2020-04-16 2021-10-21 浜松ホトニクス株式会社 Radiographic image processing method, trained model, radiographic image processing module, radiographic image processing program, and radiographic image processing system
WO2021210617A1 (en) * 2020-04-16 2021-10-21 浜松ホトニクス株式会社 Radiography method, trained model, radiography module, radiography program, radiography system, and machine learning method
WO2021225016A1 (en) * 2020-05-08 2021-11-11 ダイキン工業株式会社 Appearance inspection system
JP2022015575A (en) * 2020-07-09 2022-01-21 コニカミノルタ株式会社 Anomaly detection system, learning apparatus, anomaly detection program, learning program, anomaly detection method, and learning method
JP2022015573A (en) * 2020-07-09 2022-01-21 コニカミノルタ株式会社 Anomaly detection system, learning apparatus, anomaly detection program, learning program, anomaly detection method, and learning method of learning method calculation apparatus
JPWO2022024245A1 (en) * 2020-07-29 2022-02-03
WO2022074897A1 (en) * 2020-10-07 2022-04-14 Ckd株式会社 Solder printing inspection device
KR102426829B1 (en) * 2021-12-16 2022-07-29 주식회사 인터엑스 Apparatus and method for product surface inspection based Artificial Intelligence
WO2022201451A1 (en) * 2021-03-25 2022-09-29 株式会社日立国際電気 Detection device and detection method
WO2022201968A1 (en) * 2021-03-26 2022-09-29 コニカミノルタ株式会社 Information processing device, control program, and control method
WO2022201973A1 (en) * 2021-03-25 2022-09-29 ソニーセミコンダクタソリューションズ株式会社 Information processing system and learning model generation method
WO2022230338A1 (en) * 2021-04-27 2022-11-03 株式会社日立ハイテク System for detecting defect and computer-readable medium
KR20220159672A (en) * 2021-05-26 2022-12-05 서울대학교산학협력단 Anormaly data detection apparatus and method based on auto encoder
WO2022254747A1 (en) * 2021-06-03 2022-12-08 三菱電機株式会社 Appearance inspection device, appearance inspection method, learning device, and inference device
WO2023276595A1 (en) * 2021-06-29 2023-01-05 ＬｅａｐＭｉｎｄ株式会社 Abnormality display device, abnormality display program, abnormality display system, and abnormality display method
WO2023281890A1 (en) * 2021-07-05 2023-01-12 浜松ホトニクス株式会社 Radiographic image acquiring device, radiographic image acquiring system, and radiographic image acquisition method
JP2023005359A (en) * 2021-06-29 2023-01-18 株式会社かんでんエンジニアリング Overhead wire damage determination system
WO2023195107A1 (en) * 2022-04-06 2023-10-12 日本電気株式会社 Object evaluation device, object evaluation method, and recording medium
JP2023148422A (en) * 2022-03-30 2023-10-13 本田技研工業株式会社 Inspection device
WO2023243202A1 (en) * 2022-06-16 2023-12-21 株式会社日立製作所 Image generation method and external appearance inspection device

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH05128250A (en) * 1991-11-08 1993-05-25 Toshiba Corp Picture recognizing device
JP2018005773A (en) * 2016-07-07 2018-01-11 株式会社リコー Abnormality determination device and abnormality determination method
JP2018112550A (en) * 2017-01-12 2018-07-19 清華大学Ｔｓｉｎｇｈｕａ Ｕｎｉｖｅｒｓｉｔｙ Inspection apparatus and method for detecting firearms

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10346740B2 (en) * 2016-06-01 2019-07-09 Kla-Tencor Corp. Systems and methods incorporating a neural network and a forward physical model for semiconductor applications

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH05128250A (en) * 1991-11-08 1993-05-25 Toshiba Corp Picture recognizing device
JP2018005773A (en) * 2016-07-07 2018-01-11 株式会社リコー Abnormality determination device and abnormality determination method
JP2018112550A (en) * 2017-01-12 2018-07-19 清華大学Ｔｓｉｎｇｈｕａ Ｕｎｉｖｅｒｓｉｔｙ Inspection apparatus and method for detecting firearms

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
KASHARA RYOSUKE ET AL.: "Frozen road surface detection by abnormality detection using polarization camera and deep learning", THE 22ND SYMPOSIUM ON SENSING VIA IMAGE INFORMATION SII2016, vol. 2, no. 22, 8 June 2016 (2016-06-08), pages 1 - 4 *
LU JIAMMING ET AL.: "A method of Noise removal for color degrade images using neural networks", PROCEEDINGS OF IEICE, vol. J86A, no. 5, 1 May 2003 (2003-05-01), pages 529 - 539 *

Cited By (40)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021106566A1 (en) * 2019-11-25 2021-06-03 オムロン株式会社 Image inspection device
JP2021140739A (en) * 2020-02-28 2021-09-16 株式会社Pros Cons Program, method for creating learned model, information processing method and information processing device
JP7289427B2 (en) 2020-02-28 2023-06-12 株式会社Pros Cons Program, information processing method and information processing apparatus
WO2021210612A1 (en) * 2020-04-16 2021-10-21 浜松ホトニクス株式会社 Radiographic image acquiring device, radiographic image acquiring system, and radiographic image acquisition method
WO2021210618A1 (en) * 2020-04-16 2021-10-21 浜松ホトニクス株式会社 Radiographic image processing method, trained model, radiographic image processing module, radiographic image processing program, and radiographic image processing system
WO2021210617A1 (en) * 2020-04-16 2021-10-21 浜松ホトニクス株式会社 Radiography method, trained model, radiography module, radiography program, radiography system, and machine learning method
WO2021225016A1 (en) * 2020-05-08 2021-11-11 ダイキン工業株式会社 Appearance inspection system
JP2021177154A (en) * 2020-05-08 2021-11-11 ダイキン工業株式会社 Appearance inspection system
JP7410402B2 (en) 2020-05-08 2024-01-10 ダイキン工業株式会社 Visual inspection system
JP2022015575A (en) * 2020-07-09 2022-01-21 コニカミノルタ株式会社 Anomaly detection system, learning apparatus, anomaly detection program, learning program, anomaly detection method, and learning method
JP2022015573A (en) * 2020-07-09 2022-01-21 コニカミノルタ株式会社 Anomaly detection system, learning apparatus, anomaly detection program, learning program, anomaly detection method, and learning method of learning method calculation apparatus
JP7459696B2 (en) 2020-07-09 2024-04-02 コニカミノルタ株式会社 Anomaly detection system, learning device, anomaly detection program, learning program, anomaly detection method, and learning method Learning method for computing device
JP7459697B2 (en) 2020-07-09 2024-04-02 コニカミノルタ株式会社 Anomaly detection system, learning device, anomaly detection program, learning program, anomaly detection method, and learning method
JPWO2022024245A1 (en) * 2020-07-29 2022-02-03
WO2022024245A1 (en) * 2020-07-29 2022-02-03 オリンパス株式会社 Inspection system, learning method, inspection method, and storage medium
WO2022074897A1 (en) * 2020-10-07 2022-04-14 Ckd株式会社 Solder printing inspection device
JP7157112B2 (en) 2020-10-07 2022-10-19 Ckd株式会社 Solder print inspection device
JP2022061546A (en) * 2020-10-07 2022-04-19 Ckd株式会社 Solder print inspection device
TWI807397B (en) * 2020-10-07 2023-07-01 日商Ｃｋｄ股份有限公司 Solder printing inspection device
JP2021157847A (en) * 2020-11-16 2021-10-07 ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッドBeijing Baidu Netcom Science Technology Co., Ltd. Method, apparatus, device, and readable storage medium for recognizing abnormal license plate
JPWO2022201451A1 (en) * 2021-03-25 2022-09-29
WO2022201973A1 (en) * 2021-03-25 2022-09-29 ソニーセミコンダクタソリューションズ株式会社 Information processing system and learning model generation method
WO2022201451A1 (en) * 2021-03-25 2022-09-29 株式会社日立国際電気 Detection device and detection method
JP7436752B2 (en) 2021-03-25 2024-02-22 株式会社日立国際電気 Detection device and detection method
WO2022201968A1 (en) * 2021-03-26 2022-09-29 コニカミノルタ株式会社 Information processing device, control program, and control method
WO2022230338A1 (en) * 2021-04-27 2022-11-03 株式会社日立ハイテク System for detecting defect and computer-readable medium
KR20230153453A (en) 2021-04-27 2023-11-06 주식회사 히타치하이테크 Defect detection systems and computer-readable media
KR20220159672A (en) * 2021-05-26 2022-12-05 서울대학교산학협력단 Anormaly data detection apparatus and method based on auto encoder
KR102653349B1 (en) 2021-05-26 2024-03-29 서울대학교산학협력단 Anormaly data detection apparatus and method based on auto encoder
WO2022254747A1 (en) * 2021-06-03 2022-12-08 三菱電機株式会社 Appearance inspection device, appearance inspection method, learning device, and inference device
JP7332654B2 (en) 2021-06-29 2023-08-23 株式会社かんでんエンジニアリング Overhead line damage assessment system
WO2023276595A1 (en) * 2021-06-29 2023-01-05 ＬｅａｐＭｉｎｄ株式会社 Abnormality display device, abnormality display program, abnormality display system, and abnormality display method
JP2023005359A (en) * 2021-06-29 2023-01-18 株式会社かんでんエンジニアリング Overhead wire damage determination system
WO2023281890A1 (en) * 2021-07-05 2023-01-12 浜松ホトニクス株式会社 Radiographic image acquiring device, radiographic image acquiring system, and radiographic image acquisition method
WO2023113274A1 (en) * 2021-12-16 2023-06-22 주식회사 인터엑스 Ai-based product surface inspection device and method
KR102426829B1 (en) * 2021-12-16 2022-07-29 주식회사 인터엑스 Apparatus and method for product surface inspection based Artificial Intelligence
JP2023148422A (en) * 2022-03-30 2023-10-13 本田技研工業株式会社 Inspection device
JP7439157B2 (en) 2022-03-30 2024-02-27 本田技研工業株式会社 Inspection equipment
WO2023195107A1 (en) * 2022-04-06 2023-10-12 日本電気株式会社 Object evaluation device, object evaluation method, and recording medium
WO2023243202A1 (en) * 2022-06-16 2023-12-21 株式会社日立製作所 Image generation method and external appearance inspection device

Also Published As

Publication number Publication date
JPWO2020031984A1 (en) 2021-08-10

Similar Documents

Publication Publication Date Title
WO2020031984A1 (en) Component inspection method and inspection system
JP6879431B2 (en) Image processing equipment, image processing method and image processing program
CN110619618A (en) Surface defect detection method and device and electronic equipment
CN107316047A (en) Image processing apparatus, image processing method and storage medium
CN109685097B (en) Image detection method and device based on GAN
JP7435303B2 (en) Inspection device, unit selection device, inspection method, and inspection program
JP6908019B2 (en) Image generator and visual inspection device
KR102559021B1 (en) Apparatus and method for generating a defect image
CN113807378A (en) Training data increment method, electronic device and computer readable recording medium
CN115830004A (en) Surface defect detection method, device, computer equipment and storage medium
WO2022229976A1 (en) Anomaly detection using a convolutional neural network and feature based memories
JP7459697B2 (en) Anomaly detection system, learning device, anomaly detection program, learning program, anomaly detection method, and learning method
CN117274258A (en) Method, system, equipment and storage medium for detecting defects of main board image
US20240005477A1 (en) Index selection device, information processing device, information processing system, inspection device, inspection system, index selection method, and index selection program
WO2016092783A1 (en) Information processing apparatus, method for processing information, discriminator generating apparatus, method for generating discriminator, and program
JP7243801B2 (en) Neural network training method and defect detection method, device and storage medium
KR20230036650A (en) Defect detection method and system based on image patch
García et al. A configuration approach for convolutional neural networks used for defect detection on surfaces
JP2018116672A (en) Information processing apparatus, information processing method, and program
JP7070308B2 (en) Estimator generator, inspection device, estimator generator method, and estimator generator
Kountchev et al. Automatic detection of welding defects
WO2023224041A1 (en) Inspection device and learning method
Oz et al. A nested autoencoder approach to automated defect inspection on textured surfaces
JP7424572B1 (en) Model generation method, model generation system, model generation program, foreign object detection method, foreign object detection system, foreign object detection program and inference model
JP7446697B2 (en) Teacher data creation method and creation device

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19847218

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2020535771

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19847218

Country of ref document: EP

Kind code of ref document: A1